def _populate_policy_info(self, arm_observations, chosen_actions,
rewards_for_argmax, est_rewards):
if self._accepts_per_arm_features:
# Saving the features for the chosen action to the policy_info.
chosen_arm_features = tf.gather(params=arm_observations,
indices=chosen_actions,
batch_dims=1)
policy_info = policy_utilities.PerArmPolicyInfo(
predicted_rewards_sampled=(
rewards_for_argmax
if policy_utilities.InfoFields.PREDICTED_REWARDS_SAMPLED
in self._emit_policy_info else ()),
predicted_rewards_mean=(
tf.stack(est_rewards, axis=-1)
if policy_utilities.InfoFields.PREDICTED_REWARDS_MEAN
in self._emit_policy_info else ()),
chosen_arm_features=chosen_arm_features)
else:
policy_info = policy_utilities.PolicyInfo(
predicted_rewards_sampled=(
rewards_for_argmax
if policy_utilities.InfoFields.PREDICTED_REWARDS_SAMPLED
in self._emit_policy_info else ()),
predicted_rewards_mean=(
tf.stack(est_rewards, axis=-1)
if policy_utilities.InfoFields.PREDICTED_REWARDS_MEAN
in self._emit_policy_info else ()))
return policy_info
def testTrainPerArmAgent(self):
obs_spec = bandit_spec_utils.create_per_arm_observation_spec(2, 3, 4)
time_step_spec = ts.time_step_spec(obs_spec)
reward_net = (
global_and_arm_feature_network.create_feed_forward_per_arm_network(
obs_spec, (4, 3), (3, 4), (4, 2)))
optimizer = tf.compat.v1.train.GradientDescentOptimizer(
learning_rate=0.1)
agent = greedy_agent.GreedyRewardPredictionAgent(
time_step_spec,
self._action_spec,
reward_network=reward_net,
accepts_per_arm_features=True,
optimizer=optimizer)
observations = {
bandit_spec_utils.GLOBAL_FEATURE_KEY:
tf.constant([[1, 2], [3, 4]], dtype=tf.float32),
bandit_spec_utils.PER_ARM_FEATURE_KEY:
tf.cast(tf.reshape(tf.range(24), shape=[2, 4, 3]),
dtype=tf.float32)
}
actions = np.array([0, 3], dtype=np.int32)
rewards = np.array([0.5, 3.0], dtype=np.float32)
initial_step, final_step = _get_initial_and_final_steps(
observations, rewards)
action_step = policy_step.PolicyStep(
action=tf.convert_to_tensor(actions),
info=policy_utilities.PerArmPolicyInfo(
chosen_arm_features=np.array([[1, 2, 3], [3, 2, 1]],
dtype=np.float32)))
experience = _get_experience(initial_step, action_step, final_step)
agent.train(experience, None)
self.evaluate(tf.compat.v1.initialize_all_variables())
def _populate_policy_info_spec(self, observation_spec,
observation_and_action_constraint_splitter):
predicted_rewards_mean = ()
if (policy_utilities.InfoFields.PREDICTED_REWARDS_MEAN
in self._emit_policy_info):
predicted_rewards_mean = tensor_spec.TensorSpec(
[self._num_actions], dtype=self._dtype)
predicted_rewards_sampled = ()
if (policy_utilities.InfoFields.PREDICTED_REWARDS_SAMPLED
in self._emit_policy_info):
predicted_rewards_sampled = tensor_spec.TensorSpec(
[self._num_actions], dtype=self._dtype)
if self._accepts_per_arm_features:
# The features for the chosen arm is saved to policy_info.
chosen_arm_features_info = (
policy_utilities.create_chosen_arm_features_info_spec(
observation_spec))
info_spec = policy_utilities.PerArmPolicyInfo(
predicted_rewards_mean=predicted_rewards_mean,
predicted_rewards_sampled=predicted_rewards_sampled,
chosen_arm_features=chosen_arm_features_info)
else:
info_spec = policy_utilities.PolicyInfo(
predicted_rewards_mean=predicted_rewards_mean,
predicted_rewards_sampled=predicted_rewards_sampled)
return info_spec
def _populate_policy_info_spec(self, context_spec):
predicted_rewards_mean = ()
if (policy_utilities.InfoFields.PREDICTED_REWARDS_MEAN
in self._emit_policy_info):
predicted_rewards_mean = tensor_spec.TensorSpec(
[self._num_actions], dtype=self._dtype)
predicted_rewards_sampled = ()
if (policy_utilities.InfoFields.PREDICTED_REWARDS_SAMPLED
in self._emit_policy_info):
predicted_rewards_sampled = tensor_spec.TensorSpec(
[self._num_actions], dtype=self._dtype)
if self._accepts_per_arm_features:
# The features for the chosen arm is saved to policy_info.
arm_spec = context_spec[bandit_spec_utils.PER_ARM_FEATURE_KEY]
chosen_arm_features_info = tensor_spec.TensorSpec(
dtype=arm_spec.dtype,
shape=arm_spec.shape[1:],
name='chosen_arm_features')
info_spec = policy_utilities.PerArmPolicyInfo(
predicted_rewards_mean=predicted_rewards_mean,
predicted_rewards_sampled=predicted_rewards_sampled,
chosen_arm_features=chosen_arm_features_info)
else:
info_spec = policy_utilities.PolicyInfo(
predicted_rewards_mean=predicted_rewards_mean,
predicted_rewards_sampled=predicted_rewards_sampled)
return info_spec
def testTrainPerArmAgentVariableActions(self):
num_actions = 5
obs_spec = bandit_spec_utils.create_per_arm_observation_spec(
2, 3, num_actions, add_num_actions_feature=True)
time_step_spec = time_step.time_step_spec(obs_spec)
action_spec = tensor_spec.BoundedTensorSpec(
dtype=tf.int32, shape=(), minimum=0, maximum=num_actions - 1)
encoding_dim = 10
encoder = (
global_and_arm_feature_network.create_feed_forward_common_tower_network(
obs_spec, (4, 3), (3, 4), (4, 2), encoding_dim))
agent = neural_linucb_agent.NeuralLinUCBAgent(
time_step_spec=time_step_spec,
action_spec=action_spec,
encoding_network=encoder,
encoding_network_num_train_steps=10,
encoding_dim=encoding_dim,
accepts_per_arm_features=True,
optimizer=tf.compat.v1.train.AdamOptimizer(learning_rate=0.001))
observations = {
bandit_spec_utils.GLOBAL_FEATURE_KEY:
tf.constant([[1, 2], [3, 4]], dtype=tf.float32),
bandit_spec_utils.PER_ARM_FEATURE_KEY:
tf.cast(
tf.reshape(tf.range(30), shape=[2, 5, 3]), dtype=tf.float32),
bandit_spec_utils.NUM_ACTIONS_FEATURE_KEY:
tf.constant([3, 4], dtype=tf.int32)
}
actions = np.array([0, 3], dtype=np.int32)
rewards = np.array([0.5, 3.0], dtype=np.float32)
initial_step = time_step.TimeStep(
tf.constant(
time_step.StepType.FIRST,
dtype=tf.int32,
shape=[2],
name='step_type'),
tf.constant(0.0, dtype=tf.float32, shape=[2], name='reward'),
tf.constant(1.0, dtype=tf.float32, shape=[2], name='discount'),
observations)
final_step = time_step.TimeStep(
tf.constant(
time_step.StepType.LAST,
dtype=tf.int32,
shape=[2],
name='step_type'),
tf.constant(rewards, dtype=tf.float32, name='reward'),
tf.constant(1.0, dtype=tf.float32, shape=[2], name='discount'),
observations)
action_step = policy_step.PolicyStep(
action=tf.convert_to_tensor(actions),
info=policy_utilities.PerArmPolicyInfo(
chosen_arm_features=np.array([[1, 2, 3], [3, 2, 1]],
dtype=np.float32)))
experience = _get_experience(initial_step, action_step, final_step)
loss_info, _ = agent.train(experience, None)
self.evaluate(tf.compat.v1.initialize_all_variables())
loss_value = self.evaluate(loss_info)
self.assertGreater(loss_value, 0.0)
def testLinearAgentUpdatePerArmFeatures(self,
batch_size,
context_dim,
exploration_policy,
dtype,
use_eigendecomp=False):
"""Check that the agent updates for specified actions and rewards."""
# Construct a `Trajectory` for the given action, observation, reward.
num_actions = 5
global_context_dim = context_dim
arm_context_dim = 3
initial_step, final_step = (
_get_initial_and_final_steps_with_per_arm_features(
batch_size, global_context_dim, num_actions, arm_context_dim))
action = np.random.randint(num_actions, size=batch_size, dtype=np.int32)
action_step = policy_step.PolicyStep(
action=tf.convert_to_tensor(action),
info=policy_utilities.PerArmPolicyInfo(
chosen_arm_features=np.arange(
batch_size * arm_context_dim, dtype=np.float32).reshape(
[batch_size, arm_context_dim])))
experience = _get_experience(initial_step, action_step, final_step)
# Construct an agent and perform the update.
observation_spec = bandit_spec_utils.create_per_arm_observation_spec(
context_dim, arm_context_dim, num_actions)
time_step_spec = time_step.time_step_spec(observation_spec)
action_spec = tensor_spec.BoundedTensorSpec(
dtype=tf.int32, shape=(), minimum=0, maximum=num_actions - 1)
agent = linear_agent.LinearBanditAgent(
exploration_policy=exploration_policy,
time_step_spec=time_step_spec,
action_spec=action_spec,
use_eigendecomp=use_eigendecomp,
accepts_per_arm_features=True,
dtype=dtype)
self.evaluate(agent.initialize())
loss_info = agent.train(experience)
self.evaluate(loss_info)
final_a = self.evaluate(agent.cov_matrix)
final_b = self.evaluate(agent.data_vector)
# Compute the expected updated estimates.
global_observation = experience.observation[
bandit_spec_utils.GLOBAL_FEATURE_KEY]
arm_observation = experience.policy_info.chosen_arm_features
overall_observation = tf.squeeze(
tf.concat([global_observation, arm_observation], axis=-1), axis=1)
rewards = tf.squeeze(experience.reward, axis=1)
expected_a_new = tf.matmul(
overall_observation, overall_observation, transpose_a=True)
expected_b_new = bandit_utils.sum_reward_weighted_observations(
rewards, overall_observation)
self.assertAllClose(expected_a_new, final_a[0])
self.assertAllClose(expected_b_new, final_b[0])
def testNumActions(self, dtype):
if not dtype.is_integer:
self.skipTest('testNumActions only applies to integer dtypes')
batch_size = 1000
# Create action spec, time_step and spec with max_num_arms = 4.
action_spec = tensor_spec.BoundedTensorSpec((), dtype, 0, 3)
time_step_spec, time_step_1 = self.create_time_step(
use_per_arm_features=True, num_arms=2)
_, time_step_2 = self.create_time_step(use_per_arm_features=True,
num_arms=3)
# First half of time_step batch will have num_action = 2 and second
# half will have num_actions = 3.
half_batch_size = int(batch_size / 2)
time_step = nest_utils.stack_nested_tensors(
[time_step_1] * half_batch_size + [time_step_2] * half_batch_size)
# The features for the chosen arm is saved to policy_info.
chosen_arm_features_info = (
policy_utilities.create_chosen_arm_features_info_spec(
time_step_spec.observation))
info_spec = policy_utilities.PerArmPolicyInfo(
chosen_arm_features=chosen_arm_features_info)
policy = random_tf_policy.RandomTFPolicy(time_step_spec=time_step_spec,
action_spec=action_spec,
info_spec=info_spec,
accepts_per_arm_features=True,
emit_log_probability=True)
action_step = policy.action(time_step)
tf.nest.assert_same_structure(action_spec, action_step.action)
# Sample from the policy 1000 times, and ensure that actions considered
# invalid according to the mask are never chosen.
step = self.evaluate(action_step)
action_ = step.action
self.assertTrue(np.all(action_ >= 0))
self.assertTrue(np.all(action_[:half_batch_size] < 2))
self.assertTrue(np.all(action_[half_batch_size:] < 3))
# With num_action valid actions, probabilities should be 1/num_actions.
self.assertAllClose(
step.info.log_probability[:half_batch_size],
tf.constant(np.log(1. / 2), shape=[half_batch_size]))
self.assertAllClose(
step.info.log_probability[half_batch_size:],
tf.constant(np.log(1. / 3), shape=[half_batch_size]))
def testTrainPerArmAgent(self):
obs_spec = bandit_spec_utils.create_per_arm_observation_spec(
2, 3, 4, add_num_actions_feature=True)
time_step_spec = ts.time_step_spec(observation_spec=obs_spec,
reward_spec=tensor_spec.TensorSpec(
[3], tf.float32))
action_spec = tensor_spec.BoundedTensorSpec((), tf.int32, 0, 3)
networks_and_loss_fns = [
(global_and_arm_feature_network.
create_feed_forward_common_tower_network(obs_spec, (4, 3), (3, 4),
(4, 2)),
tf.compat.v1.losses.mean_squared_error) for _ in range(3)
]
optimizer = tf.compat.v1.train.GradientDescentOptimizer(
learning_rate=0.01)
agent = greedy_multi_objective_agent.GreedyMultiObjectiveNeuralAgent(
time_step_spec,
action_spec,
self._scalarizer,
objective_network_and_loss_fn_sequence=networks_and_loss_fns,
accepts_per_arm_features=True,
optimizer=optimizer)
observations = {
bandit_spec_utils.GLOBAL_FEATURE_KEY:
tf.constant([[1, 2], [3, 4]], dtype=tf.float32),
bandit_spec_utils.PER_ARM_FEATURE_KEY:
tf.cast(tf.reshape(tf.range(24), shape=[2, 4, 3]),
dtype=tf.float32),
bandit_spec_utils.NUM_ACTIONS_FEATURE_KEY:
tf.ones([2], dtype=tf.int32)
}
actions = np.array([0, 3], dtype=np.int32)
objectives = np.array([[1, 2, 3], [6, 5, 4]], dtype=np.float32)
initial_step, final_step = _get_initial_and_final_steps(
observations, objectives)
action_step = policy_step.PolicyStep(
action=tf.convert_to_tensor(actions),
info=policy_utilities.PerArmPolicyInfo(
chosen_arm_features=np.array([[1, 2, 3], [3, 2, 1]],
dtype=np.float32)))
experience = _get_experience(initial_step, action_step, final_step)
agent.train(experience, None)
self.evaluate(tf.compat.v1.initialize_all_variables())
def testProcessExperiencePerArmFeaturesWithMask(self):
mask_spec = tensor_spec.BoundedTensorSpec(shape=(5, ),
minimum=0,
maximum=1,
dtype=tf.int32)
observation_spec = ({
'global':
tf.TensorSpec(shape=(4, ), dtype=tf.float32),
'per_arm': {
'f1': tf.TensorSpec(shape=(5, ), dtype=tf.string),
'f2': tf.TensorSpec(shape=(5, 2), dtype=tf.int32)
}
}, mask_spec)
time_step_spec = time_step.time_step_spec(observation_spec)
policy_info_spec = policy_utilities.PerArmPolicyInfo(
chosen_arm_features={
'f1': tf.TensorSpec(shape=(), dtype=tf.string),
'f2': tf.TensorSpec(shape=(2, ), dtype=tf.int32)
})
training_data_spec = trajectory.Trajectory(
step_type=time_step_spec.step_type,
observation=time_step_spec.observation,
action=tensor_spec.BoundedTensorSpec(shape=(),
minimum=0,
maximum=4,
dtype=tf.int32),
policy_info=policy_info_spec,
next_step_type=time_step_spec.step_type,
reward=tensor_spec.BoundedTensorSpec(shape=(),
minimum=0,
maximum=2,
dtype=tf.float32),
discount=time_step_spec.discount)
experience = tensor_spec.sample_spec_nest(training_data_spec,
outer_dims=(7, 2))
observation, action, reward = utils.process_experience_for_neural_agents(
experience, lambda x: (x[0], x[1]), True, training_data_spec)
self.assertEqual(
observation['per_arm']['f1'][0],
experience.policy_info.chosen_arm_features['f1'][0, 0])
self.assertAllEqual(action, tf.zeros(14, dtype=tf.int32))
self.assertEqual(reward[0], experience.reward[0, 0])
def testTrainPerArmAgentWithMask(self):
num_actions = 4
obs_spec = bandit_spec_utils.create_per_arm_observation_spec(
2, 3, num_actions)
mask_obs_spec = (obs_spec,
tensor_spec.BoundedTensorSpec(shape=[num_actions],
minimum=0,
maximum=1,
dtype=tf.float32))
time_step_spec = ts.time_step_spec(mask_obs_spec)
reward_net = (global_and_arm_feature_network.
create_feed_forward_common_tower_network(
obs_spec, (4, 3), (3, 4), (4, 2)))
optimizer = tf.compat.v1.train.GradientDescentOptimizer(
learning_rate=0.1)
agent = greedy_agent.GreedyRewardPredictionAgent(
time_step_spec,
self._action_spec,
reward_network=reward_net,
observation_and_action_constraint_splitter=lambda x: [x[0], x[1]],
accepts_per_arm_features=True,
optimizer=optimizer)
observations = ({
bandit_spec_utils.GLOBAL_FEATURE_KEY:
tf.constant([[1, 2], [3, 4]], dtype=tf.float32),
bandit_spec_utils.PER_ARM_FEATURE_KEY:
tf.cast(tf.reshape(tf.range(24), shape=[2, 4, 3]),
dtype=tf.float32)
}, tf.ones([2, num_actions]))
actions = np.array([0, 3], dtype=np.int32)
rewards = np.array([0.5, 3.0], dtype=np.float32)
initial_step, final_step = _get_initial_and_final_steps_with_action_mask(
observations, rewards)
action_step = policy_step.PolicyStep(
action=tf.convert_to_tensor(actions),
info=policy_utilities.PerArmPolicyInfo(
chosen_arm_features=np.array([[1, 2, 3], [3, 2, 1]],
dtype=np.float32)))
experience = _get_experience(initial_step, action_step, final_step)
agent.train(experience, None)
self.evaluate(tf.compat.v1.initialize_all_variables())
def __init__(self,
time_step_spec: types.TimeStep,
action_spec: types.NestedTensorSpec,
reward_network: types.Network,
observation_and_action_constraint_splitter: Optional[
types.Splitter] = None,
accepts_per_arm_features: bool = False,
constraints: Tuple[constr.NeuralConstraint, ...] = (),
emit_policy_info: Tuple[Text, ...] = (),
name: Optional[Text] = None):
"""Builds a GreedyRewardPredictionPolicy given a reward tf_agents.Network.
This policy takes a tf_agents.Network predicting rewards and generates the
action corresponding to the largest predicted reward.
Args:
time_step_spec: A `TimeStep` spec of the expected time_steps.
action_spec: A nest of BoundedTensorSpec representing the actions.
reward_network: An instance of a `tf_agents.network.Network`,
callable via `network(observation, step_type) -> (output, final_state)`.
observation_and_action_constraint_splitter: A function used for masking
valid/invalid actions with each state of the environment. The function
takes in a full observation and returns a tuple consisting of 1) the
part of the observation intended as input to the network and 2) the
mask. The mask should be a 0-1 `Tensor` of shape
`[batch_size, num_actions]`. This function should also work with a
`TensorSpec` as input, and should output `TensorSpec` objects for the
observation and mask.
accepts_per_arm_features: (bool) Whether the policy accepts per-arm
features.
constraints: iterable of constraints objects that are instances of
`tf_agents.bandits.agents.NeuralConstraint`.
emit_policy_info: (tuple of strings) what side information we want to get
as part of the policy info. Allowed values can be found in
`policy_utilities.PolicyInfo`.
name: The name of this policy. All variables in this module will fall
under that name. Defaults to the class name.
Raises:
NotImplementedError: If `action_spec` contains more than one
`BoundedTensorSpec` or the `BoundedTensorSpec` is not valid.
"""
policy_utilities.check_no_mask_with_arm_features(
accepts_per_arm_features,
observation_and_action_constraint_splitter)
flat_action_spec = tf.nest.flatten(action_spec)
if len(flat_action_spec) > 1:
raise NotImplementedError(
'action_spec can only contain a single BoundedTensorSpec.')
action_spec = flat_action_spec[0]
if (not tensor_spec.is_bounded(action_spec)
or not tensor_spec.is_discrete(action_spec)
or action_spec.shape.rank > 1
or action_spec.shape.num_elements() != 1):
raise NotImplementedError(
'action_spec must be a BoundedTensorSpec of type int32 and shape (). '
'Found {}.'.format(action_spec))
self._expected_num_actions = action_spec.maximum - action_spec.minimum + 1
self._action_offset = action_spec.minimum
reward_network.create_variables()
self._reward_network = reward_network
self._constraints = constraints
self._emit_policy_info = emit_policy_info
predicted_rewards_mean = ()
if policy_utilities.InfoFields.PREDICTED_REWARDS_MEAN in emit_policy_info:
predicted_rewards_mean = tensor_spec.TensorSpec(
[self._expected_num_actions])
bandit_policy_type = ()
if policy_utilities.InfoFields.BANDIT_POLICY_TYPE in emit_policy_info:
bandit_policy_type = (
policy_utilities.create_bandit_policy_type_tensor_spec(
shape=[1]))
if accepts_per_arm_features:
# The features for the chosen arm is saved to policy_info.
chosen_arm_features_info = (
policy_utilities.create_chosen_arm_features_info_spec(
time_step_spec.observation))
info_spec = policy_utilities.PerArmPolicyInfo(
predicted_rewards_mean=predicted_rewards_mean,
bandit_policy_type=bandit_policy_type,
chosen_arm_features=chosen_arm_features_info)
else:
info_spec = policy_utilities.PolicyInfo(
predicted_rewards_mean=predicted_rewards_mean,
bandit_policy_type=bandit_policy_type)
self._accepts_per_arm_features = accepts_per_arm_features
super(GreedyRewardPredictionPolicy,
self).__init__(time_step_spec,
action_spec,
policy_state_spec=reward_network.state_spec,
clip=False,
info_spec=info_spec,
emit_log_probability='log_probability'
in emit_policy_info,
observation_and_action_constraint_splitter=(
observation_and_action_constraint_splitter),
name=name)
def _distribution(self, time_step, policy_state):
observation = time_step.observation
if self.observation_and_action_constraint_splitter is not None:
observation, _ = self.observation_and_action_constraint_splitter(
observation)
predictions, policy_state = self._reward_network(
observation, time_step.step_type, policy_state)
batch_size = tf.shape(predictions)[0]
if isinstance(self._reward_network,
heteroscedastic_q_network.HeteroscedasticQNetwork):
predicted_reward_values = predictions.q_value_logits
else:
predicted_reward_values = predictions
predicted_reward_values.shape.with_rank_at_least(2)
predicted_reward_values.shape.with_rank_at_most(3)
if predicted_reward_values.shape[
-1] is not None and predicted_reward_values.shape[
-1] != self._expected_num_actions:
raise ValueError(
'The number of actions ({}) does not match the reward_network output'
' size ({}).'.format(self._expected_num_actions,
predicted_reward_values.shape[1]))
mask = constr.construct_mask_from_multiple_sources(
time_step.observation,
self._observation_and_action_constraint_splitter,
self._constraints, self._expected_num_actions)
# Argmax.
if mask is not None:
actions = policy_utilities.masked_argmax(
predicted_reward_values,
mask,
output_type=self.action_spec.dtype)
else:
actions = tf.argmax(predicted_reward_values,
axis=-1,
output_type=self.action_spec.dtype)
actions += self._action_offset
bandit_policy_values = tf.fill(
[batch_size, 1], policy_utilities.BanditPolicyType.GREEDY)
if self._accepts_per_arm_features:
# Saving the features for the chosen action to the policy_info.
def gather_observation(obs):
return tf.gather(params=obs, indices=actions, batch_dims=1)
chosen_arm_features = tf.nest.map_structure(
gather_observation,
observation[bandit_spec_utils.PER_ARM_FEATURE_KEY])
policy_info = policy_utilities.PerArmPolicyInfo(
log_probability=tf.zeros([batch_size], tf.float32)
if policy_utilities.InfoFields.LOG_PROBABILITY
in self._emit_policy_info else (),
predicted_rewards_mean=(
predicted_reward_values
if policy_utilities.InfoFields.PREDICTED_REWARDS_MEAN
in self._emit_policy_info else ()),
bandit_policy_type=(
bandit_policy_values
if policy_utilities.InfoFields.BANDIT_POLICY_TYPE
in self._emit_policy_info else ()),
chosen_arm_features=chosen_arm_features)
else:
policy_info = policy_utilities.PolicyInfo(
log_probability=tf.zeros([batch_size], tf.float32)
if policy_utilities.InfoFields.LOG_PROBABILITY
in self._emit_policy_info else (),
predicted_rewards_mean=(
predicted_reward_values
if policy_utilities.InfoFields.PREDICTED_REWARDS_MEAN
in self._emit_policy_info else ()),
bandit_policy_type=(
bandit_policy_values
if policy_utilities.InfoFields.BANDIT_POLICY_TYPE
in self._emit_policy_info else ()))
return policy_step.PolicyStep(
tfp.distributions.Deterministic(loc=actions), policy_state,
policy_info)
def __init__(
self,
time_step_spec: Optional[ts.TimeStep],
action_spec: Optional[NestedBoundedTensorSpec],
scalarizer: multi_objective_scalarizer.Scalarizer,
objective_networks: Sequence[Network],
observation_and_action_constraint_splitter: types.Splitter = None,
accepts_per_arm_features: bool = False,
emit_policy_info: Tuple[Text] = (),
name: Optional[Text] = None):
"""Builds a GreedyMultiObjectiveNeuralPolicy based on multiple networks.
This policy takes an iterable of `tf_agents.Network`, each responsible for
predicting a specific objective, along with a `Scalarizer` object to
generate an action by maximizing the scalarized objective, i.e., the output
of the `Scalarizer` applied to the multiple predicted objectives by the
networks.
Args:
time_step_spec: A `TimeStep` spec of the expected time_steps.
action_spec: A nest of `BoundedTensorSpec` representing the actions.
scalarizer: A
`tf_agents.bandits.multi_objective.multi_objective_scalarizer.Scalarizer`
object that implements scalarization of multiple objectives into a
single scalar reward.
objective_networks: A Sequence of `tf_agents.network.Network` objects to
be used by the policy. Each network will be called with
call(observation, step_type) and is expected to provide a prediction for
a specific objective for all actions.
observation_and_action_constraint_splitter: A function used for masking
valid/invalid actions with each state of the environment. The function
takes in a full observation and returns a tuple consisting of 1) the
part of the observation intended as input to the network and 2) the
mask. The mask should be a 0-1 `Tensor` of shape `[batch_size,
num_actions]`. This function should also work with a `TensorSpec` as
input, and should output `TensorSpec` objects for the observation and
mask.
accepts_per_arm_features: (bool) Whether the policy accepts per-arm
features.
emit_policy_info: (tuple of strings) what side information we want to get
as part of the policy info. Allowed values can be found in
`policy_utilities.PolicyInfo`.
name: The name of this policy. All variables in this module will fall
under that name. Defaults to the class name.
Raises:
NotImplementedError: If `action_spec` contains more than one
`BoundedTensorSpec` or the `BoundedTensorSpec` is not valid.
NotImplementedError: If `action_spec` is not a `BoundedTensorSpec` of type
int32 and shape ().
ValueError: If `objective_networks` has fewer than two networks.
ValueError: If `accepts_per_arm_features` is true but `time_step_spec` is
None.
"""
flat_action_spec = tf.nest.flatten(action_spec)
if len(flat_action_spec) > 1:
raise NotImplementedError(
'action_spec can only contain a single BoundedTensorSpec.')
action_spec = flat_action_spec[0]
if (not tensor_spec.is_bounded(action_spec)
or not tensor_spec.is_discrete(action_spec)
or action_spec.shape.rank > 1
or action_spec.shape.num_elements() != 1):
raise NotImplementedError(
'action_spec must be a BoundedTensorSpec of type int32 and shape (). '
'Found {}.'.format(action_spec))
self._expected_num_actions = action_spec.maximum - action_spec.minimum + 1
self._action_offset = action_spec.minimum
policy_state_spec = []
for network in objective_networks:
policy_state_spec.append(network.state_spec)
network.create_variables()
self._objective_networks = objective_networks
self._scalarizer = scalarizer
self._num_objectives = len(self._objective_networks)
if self._num_objectives < 2:
raise ValueError(
'Number of objectives should be at least two, but found to be {}'
.format(self._num_objectives))
self._emit_policy_info = emit_policy_info
predicted_rewards_mean = ()
if policy_utilities.InfoFields.PREDICTED_REWARDS_MEAN in emit_policy_info:
predicted_rewards_mean = tensor_spec.TensorSpec(
[self._num_objectives, self._expected_num_actions])
bandit_policy_type = ()
if policy_utilities.InfoFields.BANDIT_POLICY_TYPE in emit_policy_info:
bandit_policy_type = (
policy_utilities.create_bandit_policy_type_tensor_spec(
shape=[1]))
if accepts_per_arm_features:
if time_step_spec is None:
raise ValueError(
'time_step_spec should not be None for per-arm-features policies, '
'but found to be.')
# The features for the chosen arm is saved to policy_info.
chosen_arm_features_info = (
policy_utilities.create_chosen_arm_features_info_spec(
time_step_spec.observation,
observation_and_action_constraint_splitter))
info_spec = policy_utilities.PerArmPolicyInfo(
predicted_rewards_mean=predicted_rewards_mean,
bandit_policy_type=bandit_policy_type,
chosen_arm_features=chosen_arm_features_info)
else:
info_spec = policy_utilities.PolicyInfo(
predicted_rewards_mean=predicted_rewards_mean,
bandit_policy_type=bandit_policy_type)
self._accepts_per_arm_features = accepts_per_arm_features
super(GreedyMultiObjectiveNeuralPolicy,
self).__init__(time_step_spec,
action_spec,
policy_state_spec=policy_state_spec,
clip=False,
info_spec=info_spec,
emit_log_probability='log_probability'
in emit_policy_info,
observation_and_action_constraint_splitter=(
observation_and_action_constraint_splitter),
name=name)
def _distribution(
self, time_step: ts.TimeStep,
policy_state: Sequence[types.TensorSpec]
) -> policy_step.PolicyStep:
observation = time_step.observation
if self.observation_and_action_constraint_splitter is not None:
observation, _ = self.observation_and_action_constraint_splitter(
observation)
predicted_objective_values_tensor, policy_state = self._predict(
observation, time_step.step_type, policy_state)
scalarized_reward = scalarize_objectives(
predicted_objective_values_tensor, self._scalarizer)
batch_size = scalarized_reward.shape[0]
mask = policy_utilities.construct_mask_from_multiple_sources(
time_step.observation,
self._observation_and_action_constraint_splitter, (),
self._expected_num_actions)
# Argmax.
if mask is not None:
actions = policy_utilities.masked_argmax(
scalarized_reward, mask, output_type=self.action_spec.dtype)
else:
actions = tf.argmax(scalarized_reward,
axis=-1,
output_type=self.action_spec.dtype)
actions += self._action_offset
bandit_policy_values = tf.fill(
[batch_size, 1], policy_utilities.BanditPolicyType.GREEDY)
if self._accepts_per_arm_features:
# Saving the features for the chosen action to the policy_info.
def gather_observation(obs):
return tf.gather(params=obs, indices=actions, batch_dims=1)
chosen_arm_features = tf.nest.map_structure(
gather_observation,
observation[bandit_spec_utils.PER_ARM_FEATURE_KEY])
policy_info = policy_utilities.PerArmPolicyInfo(
log_probability=tf.zeros([batch_size], tf.float32)
if policy_utilities.InfoFields.LOG_PROBABILITY
in self._emit_policy_info else (),
predicted_rewards_mean=(
predicted_objective_values_tensor
if policy_utilities.InfoFields.PREDICTED_REWARDS_MEAN
in self._emit_policy_info else ()),
bandit_policy_type=(
bandit_policy_values
if policy_utilities.InfoFields.BANDIT_POLICY_TYPE
in self._emit_policy_info else ()),
chosen_arm_features=chosen_arm_features)
else:
policy_info = policy_utilities.PolicyInfo(
log_probability=tf.zeros([batch_size], tf.float32)
if policy_utilities.InfoFields.LOG_PROBABILITY
in self._emit_policy_info else (),
predicted_rewards_mean=(
predicted_objective_values_tensor
if policy_utilities.InfoFields.PREDICTED_REWARDS_MEAN
in self._emit_policy_info else ()),
bandit_policy_type=(
bandit_policy_values
if policy_utilities.InfoFields.BANDIT_POLICY_TYPE
in self._emit_policy_info else ()))
return policy_step.PolicyStep(
tfp.distributions.Deterministic(loc=actions), policy_state,
policy_info)
def __init__(self,
encoding_network,
encoding_dim,
reward_layer,
epsilon_greedy,
actions_from_reward_layer,
cov_matrix,
data_vector,
num_samples,
time_step_spec=None,
alpha=1.0,
emit_policy_info=(),
emit_log_probability=False,
accepts_per_arm_features=False,
observation_and_action_constraint_splitter=None,
name=None):
"""Initializes `NeuralLinUCBPolicy`.
Args:
encoding_network: network that encodes the observations.
encoding_dim: (int) dimension of the encoded observations.
reward_layer: final layer that predicts the expected reward per arm. In
case the policy accepts per-arm features, the output of this layer has
to be a scalar. This is because in the per-arm case, all encoded
observations have to go through the same computation to get the reward
estimates. The `num_actions` dimension of the encoded observation is
treated as a batch dimension in the reward layer.
epsilon_greedy: (float) representing the probability of choosing a random
action instead of the greedy action.
actions_from_reward_layer: (bool) whether to get actions from the reward
layer or from LinUCB.
cov_matrix: list of the covariance matrices. There exists one covariance
matrix per arm, unless the policy accepts per-arm features, in which
case this list must have a single element.
data_vector: list of the data vectors. A data vector is a weighted sum
of the observations, where the weight is the corresponding reward. Each
arm has its own data vector, unless the policy accepts per-arm features,
in which case this list must have a single element.
num_samples: list of number of samples per arm. If the policy accepts per-
arm features, this is a single-element list counting the number of
steps.
time_step_spec: A `TimeStep` spec of the expected time_steps.
alpha: (float) non-negative weight multiplying the confidence intervals.
emit_policy_info: (tuple of strings) what side information we want to get
as part of the policy info. Allowed values can be found in
`policy_utilities.PolicyInfo`.
emit_log_probability: (bool) whether to emit log probabilities.
accepts_per_arm_features: (bool) Whether the policy accepts per-arm
features.
observation_and_action_constraint_splitter: A function used for masking
valid/invalid actions with each state of the environment. The function
takes in a full observation and returns a tuple consisting of 1) the
part of the observation intended as input to the bandit policy and 2)
the mask. The mask should be a 0-1 `Tensor` of shape
`[batch_size, num_actions]`. This function should also work with a
`TensorSpec` as input, and should output `TensorSpec` objects for the
observation and mask.
name: The name of this policy.
"""
encoding_network.create_variables()
self._encoding_network = encoding_network
self._reward_layer = reward_layer
self._encoding_dim = encoding_dim
if accepts_per_arm_features and reward_layer.units != 1:
raise ValueError(
'The output dimension of the reward layer must be 1, got'
' {}'.format(reward_layer.units))
if not isinstance(cov_matrix, (list, tuple)):
raise ValueError(
'cov_matrix must be a list of matrices (Tensors).')
self._cov_matrix = cov_matrix
if not isinstance(data_vector, (list, tuple)):
raise ValueError(
'data_vector must be a list of vectors (Tensors).')
self._data_vector = data_vector
if not isinstance(num_samples, (list, tuple)):
raise ValueError(
'num_samples must be a list of vectors (Tensors).')
self._num_samples = num_samples
self._alpha = alpha
self._actions_from_reward_layer = actions_from_reward_layer
self._epsilon_greedy = epsilon_greedy
self._dtype = self._data_vector[0].dtype
if len(cov_matrix) != len(data_vector):
raise ValueError(
'The size of list cov_matrix must match the size of '
'list data_vector. Got {} for cov_matrix and {} '
'for data_vector'.format(len(self._cov_matrix),
len((data_vector))))
if len(num_samples) != len(cov_matrix):
raise ValueError('The size of num_samples must match the size of '
'list cov_matrix. Got {} for num_samples and {} '
'for cov_matrix'.format(len(self._num_samples),
len((cov_matrix))))
self._accepts_per_arm_features = accepts_per_arm_features
if observation_and_action_constraint_splitter is not None:
context_spec, _ = observation_and_action_constraint_splitter(
time_step_spec.observation)
else:
context_spec = time_step_spec.observation
if accepts_per_arm_features:
self._num_actions = context_spec[
bandit_spec_utils.PER_ARM_FEATURE_KEY].shape.as_list()[0]
self._num_models = 1
else:
self._num_actions = len(cov_matrix)
self._num_models = self._num_actions
(self._global_context_dim,
self._arm_context_dim) = bandit_spec_utils.get_context_dims_from_spec(
context_spec, accepts_per_arm_features)
cov_matrix_dim = tf.compat.dimension_value(cov_matrix[0].shape[0])
if self._encoding_dim != cov_matrix_dim:
raise ValueError('The dimension of matrix `cov_matrix` must match '
'encoding dimension {}.'
'Got {} for `cov_matrix`.'.format(
self._encoding_dim, cov_matrix_dim))
data_vector_dim = tf.compat.dimension_value(data_vector[0].shape[0])
if self._encoding_dim != data_vector_dim:
raise ValueError(
'The dimension of vector `data_vector` must match '
'encoding dimension {}. '
'Got {} for `data_vector`.'.format(self._encoding_dim,
data_vector_dim))
action_spec = tensor_spec.BoundedTensorSpec(shape=(),
dtype=tf.int32,
minimum=0,
maximum=self._num_actions -
1,
name='action')
self._emit_policy_info = emit_policy_info
predicted_rewards_mean = ()
if policy_utilities.InfoFields.PREDICTED_REWARDS_MEAN in emit_policy_info:
predicted_rewards_mean = tensor_spec.TensorSpec(
[self._num_actions], dtype=tf.float32)
if accepts_per_arm_features:
chosen_arm_features_info = tensor_spec.TensorSpec(
dtype=tf.float32,
shape=[self._arm_context_dim],
name='chosen_arm_features')
info_spec = policy_utilities.PerArmPolicyInfo(
predicted_rewards_mean=predicted_rewards_mean,
chosen_arm_features=chosen_arm_features_info)
else:
info_spec = policy_utilities.PolicyInfo(
predicted_rewards_mean=predicted_rewards_mean)
super(NeuralLinUCBPolicy,
self).__init__(time_step_spec=time_step_spec,
action_spec=action_spec,
emit_log_probability=emit_log_probability,
observation_and_action_constraint_splitter=(
observation_and_action_constraint_splitter),
info_spec=info_spec,
name=name)
